Re: Time for NUMA, Embrace your Inner CSP
You're missing the point. Naively written anything will overwhelm underlying hardware. There's nothing magical about shared memory in a SMP-on-top-of-NUMA system that means that poorly written code won't run into the limits of the QPI / Hypertransport links between CPUs.
Synthesising SMP on top of NUMA requires a lot of traffic to flow over these links to achieve cache coherency. Ditch the SMP, and you've also ditched the cache coherency traffic on these links, meaning that there's more link time available for other traffic (such as data transfers for a CSP system). And you've got rid of a whole class of hardware flaws revealed in the article, and you have a faster system. What's not to like?
From what I hear from my local Rust enthusiast, Rust's control of memory ownership boils down to being the same as CSP. Certainly, Rust has the same concept of synchronous channels that CSP has.
One of the good things about CSP is that it makes it abundantly clear that one has written rubbish code; there's a lot of channel reads and writes littering one's code.